README.md

Warehouse Object Detection Dataset

Overview

This is a synthetic warehouse object detection dataset generated using NVIDIA Omniverse Replicator. The dataset contains high-quality RGB images with comprehensive annotations including 2D bounding boxes, instance segmentation masks, depth maps, and 3D primitive paths.

Version: 1.0.0 License: CC-BY-4.0 Total Images: 6,234 Total Annotations: 78,441 Image Resolution: 512x512 Number of Classes: 25

Dataset Statistics

Split Distribution

• Training Set: 4,363 images (70.0%)
• Validation Set: 935 images (15.0%)
• Test Set: 936 images (15.0%)

Annotation Statistics

• Average Annotations per Image: 12.6
• Total Bounding Boxes: 78,441

Top 10 Most Common Classes

1. wall: 6,234 instances in 6,234 images
2. floor: 6,232 instances in 6,232 images
3. sign: 5,945 instances in 5,945 images
4. floor_decal: 5,945 instances in 5,945 images
5. pillar: 5,845 instances in 5,845 images
6. rack: 5,824 instances in 5,824 images
7. box: 5,789 instances in 5,789 images
8. pallet: 5,743 instances in 5,743 images
9. bracket: 4,857 instances in 4,857 images
10. lamp: 4,451 instances in 4,451 images

Classes

The dataset includes 25 object classes organized into the following categories:

Container

• box: 5,789 annotations
• crate: 1,270 annotations
• barrel: 1,229 annotations
• bottle: 665 annotations
• bucket: 457 annotations

Storage

• pallet: 5,743 annotations
• rack: 5,824 annotations

Infrastructure

• bracket: 4,857 annotations
• pillar: 5,845 annotations
• emergency_board: 569 annotations

Equipment

• lamp: 4,451 annotations
• sign: 5,945 annotations
• wire: 3,985 annotations
• fuse_box: 1,154 annotations
• fire_extinguisher: 3,304 annotations
• forklift: 343 annotations
• cart: 427 annotations
• cone: 639 annotations

Markers

• floor_decal: 5,945 annotations
• barcode: 1,137 annotations
• paper_note: 2,125 annotations
• paper_shortcut: 390 annotations

Background

• wall: 6,234 annotations
• ceiling: 3,882 annotations
• floor: 6,232 annotations

Dataset Formats

This dataset is provided in three formats to support different use cases:

1. Raw Format (raw/)

Preserves all original Omniverse data:

• RGB images (PNG)
• 2D bounding boxes (NumPy .npy)
• Class labels (JSON)
• 3D primitive paths (JSON)
• Instance segmentation masks (PNG)
• Instance ID mappings (JSON)
• Depth/distance maps (NumPy .npy)

Use for: Multi-task learning, depth estimation, 3D tasks, research

2. YOLO Format (yolo/)

Standard YOLO v5/v8 format:

• Images in images/train/, images/val/, images/test/
• Labels in labels/train/, labels/val/, labels/test/
• Each label file contains: class_id x_center y_center width height (normalized 0-1)

Use for: Object detection with YOLO models (Ultralytics, Darknet)

3. COCO Format (coco/)

COCO JSON format:

• Images in images/train/, images/val/, images/test/
• Annotations in annotations/instances_{train,val,test}.json

Use for: Object detection/segmentation with detectron2, MMDetection, etc.

Directory Structure

warehouse_detection_dataset/
├── README.md                    # This file
├── data.yaml                    # YOLO configuration
├── dataset_info.json            # HuggingFace metadata
├── class_mapping.json           # Complete class information
├── raw/                         # Original Omniverse format
│   ├── train/
│   │   ├── images/
│   │   ├── annotations/
│   │   ├── segmentation/
│   │   └── depth/
│   ├── val/
│   └── test/
├── yolo/                        # YOLO format
│   ├── images/
│   │   ├── train/
│   │   ├── val/
│   │   └── test/
│   └── labels/
│       ├── train/
│       ├── val/
│       └── test/
└── coco/                        # COCO format
    ├── images/
    │   ├── train/
    │   ├── val/
    │   └── test/
    └── annotations/
        ├── instances_train.json
        ├── instances_val.json
        └── instances_test.json

Usage Examples

YOLO (Ultralytics)

from ultralytics import YOLO

# Train a model
model = YOLO('yolov8n.pt')
model.train(data='warehouse_detection_dataset/data.yaml', epochs=100)

# Validate
metrics = model.val()

# Predict
results = model('path/to/image.jpg')

COCO (detectron2)

from detectron2.data import DatasetCatalog, MetadataCatalog
from detectron2.data.datasets import load_coco_json

# Register dataset
DatasetCatalog.register(
    "warehouse_train",
    lambda: load_coco_json(
        "warehouse_detection_dataset/coco/annotations/instances_train.json",
        "warehouse_detection_dataset/coco/images/train"
    )
)

# Train your model
# ... (standard detectron2 training code)

Raw Format (Custom)

import numpy as np
import json
from PIL import Image

# Load image
img = Image.open('raw/train/images/warehouse_000001.png')

# Load bounding boxes
bboxes = np.load('raw/train/annotations/warehouse_000001_bbox.npy')

# Load labels
with open('raw/train/annotations/warehouse_000001_labels.json') as f:
    labels = json.load(f)

# Load segmentation mask
seg_mask = Image.open('raw/train/segmentation/warehouse_000001_seg.png')

# Load depth map
depth = np.load('raw/train/depth/warehouse_000001_depth.npy')

Citiation

If you use this dataset in your research, please cite:

@dataset{warehouse_detection_2025,
  title={Warehouse Object Detection Dataset},
  author={Howe, McCarthy and Phillips, Cassandra and Lee, Alfred and Sethi, Varun and Hall, Jada},
  year={2025},
  publisher={Clemson University Capstone},
  note={In partnership with Capgemini Supply Chain},
  version={1.0.0},
  license={CC-BY-4.0}
}

License

This dataset is released under the CC-BY-4.0 license.

Team & Acknowledgments

Project Context: This dataset was created as part of the Clemson University Capstone 2025 project, in partnership with Capgemini Supply Chain.

Contributors:

• McCarthy Howe (mac@machowe.com)
• Cassandra Phillips
• Alfred Lee
• Varun Sethi
• Jada Hall

Tooling: Generated using NVIDIA Omniverse Replicator.

Contact

For questions, issues, or feedback, please contact McCarthy Howe at mac@machowe.com.